A million-token window, weights included

Meta has published the weights for Muse Spark Open, which it describes as the most capable US open-weight model to date. The announcement appeared on research.meta.ai on 15 October 2026. The model is natively multimodal, with 816B total parameters and 42B active, and a context window of 1M tokens.

Two figures carry most of the weight here. The 1M-token window is a capacity statement: a single pass can take in a long document set rather than a selected excerpt of it. The open weights are a placement statement: the model can be downloaded and run outside a vendor's API. Meta pairs the second with a specific desktop configuration.

Two Mac Studios hold the whole model

Meta states that Muse Spark Open is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the latest MLX framework. Apple's unified memory architecture lets the M5 Ultra's integrated GPU address up to 512 GB per machine, so two machines can hold the entire model, and the chip's 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. Meta's description is that the pair behaves like one computer with a terabyte of unified memory.

That is Meta's account of hardware it optimised for, not an independent measurement. What it establishes is that a local option is on the table for teams that cannot route documents through a third-party API.

Meta also reports on-device agentic tests: shipping iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end to review footage frame by frame and crop clips vertically. Those are Meta's own tests.

Kernels, licence and where to get it

The open-weight release ships with auto-tunable Helion GPU kernels in SGLang, which Meta reports as 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. Meta says it evaluated the model under its Meta Superintelligence Scaling Framework, covering the most severe risks, and that it met the thresholds required for release, with additional safety post-training for the open-weight version.

Weights are downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models, with hosted inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA silicon, and can be tuned with PyTorch.

The licence is the Muse Model License, which Meta describes as carrying limited commercial terms that allow many developers to use the model for free. The terms themselves decide who counts as one of them.

For a team weighing whether a long tender package or a contract set can be read in one pass without leaving its own infrastructure, the combination to check is the 1M-token window against the two-machine setup Meta describes. Both are stated by the vendor; the licence and the hardware bill are what a deployment decision turns on.